Writing on memory, agents, and building Hebbrix
A preview of what we're writing. These posts are on the way, not published yet.
Why we modeled memory after the Ebbinghaus forgetting curve
Vector search treats every memory as equal. Human memory doesn't, and neither does ours. A close look at why cognitive science made a better foundation for AI memory than information retrieval theory.
Vector search alone isn't enough. We run five retrieval signals in parallel (semantic, BM25, graph traversal, importance, and recency), and here's how we fuse them into one result.
The teams building the best agents have moved past crafting prompts. They're engineering what reaches the model. Here's what that shift actually looks like in practice.
How we pull entities and relationships out of unstructured text, link them into a graph, and keep extraction latency sub-second per memory, all without a schema.
We built a test suite that measures retrieval accuracy, latency, and memory quality across four memory systems. Here's what we found, including where we lost.
Our six RL quality checks run after every interaction. No thumbs-up buttons, no manual feedback loops. Here's the mechanism behind automatic memory improvement.
Memory is the most human thing we're building. Cream paper, ink, and Fraunces type felt more honest than the cyan-on-black alternative. A note on the rebrand.
We took a real LangChain agent and added Hebbrix in under 10 minutes. Session transcript, timing data, and the before/after retrieval comparison included.
New posts, roughly twice a month
Engineering walkthroughs, product updates, and notes on context engineering. Nothing else clogging your inbox.